TextVerifier Looks Beyond Recognition to Check Chinese Character Rendering

Introduction

The challenge of generating Chinese text in images is shifting. It is no longer only about whether a model can produce something that people can read, but whether every character has been rendered correctly. Text-to-image systems may generate a recognizable sentence while introducing local defects: a missing stroke, an extra stroke, or a malformed radical or component. Because the resulting glyph can remain understandable, conventional OCR and general vision-language models may overlook the problem.

TextVerifier is designed for this specific gap. The paper formulates text-rendering verification as a target-conditioned task. Given a target Chinese character and an image crop, the verifier must decide whether the crop is a correct rendering of that character—not merely identify what character it resembles.

What the method changes

• From recognition to structural verification: OCR primarily measures identity and readability. TextVerifier evaluates whether the visible structure agrees with the intended character.
• An IDS-based target: Instead of placing a single character symbol in the text tower, the system uses the character’s Ideographic Description Sequence (IDS). This representation describes how the character is composed from parts, giving the verifier a more explicit structural condition.
• Quality-sensitive training pairs: Training uses 12.7 million crop-IDS pairs. It also adds label-free hard pairs containing the same character at different quality levels, sampled from the training trajectory of a text-region inpainting model. These examples encourage the model to distinguish a readable glyph from a correctly rendered one.
• Character-level evaluation: The authors introduce TextRenderBench-ZH, annotating the position, identity, and correctness of 7,358 characters in generated Chinese text. The benchmark therefore evaluates errors at the level where they actually occur.

Results and implications

TextVerifier reaches an F1 score of 0.844 on the benchmark. The comparison figures are 0.526 for OCR confidence, 0.329 for a general VLM, and 0.590 for TextPecker. The gap highlights an important distinction: recognizing a character and verifying its rendering are related but separate capabilities. Improving recognition alone does not guarantee that a model will catch a malformed component or a subtle stroke error.

The work also demonstrates a route from evaluation to correction. The verifier’s per-character scores are used as rewards in Flow-GRPO post-training. In this setup, the model can provide localized feedback during optimization, rather than treating an entire generated image as simply good or bad. The reported result is a further reduction in rendering errors.

This approach could be useful for Chinese posters, advertisements, infographics, and other visual content where a single incorrect character can change meaning or reduce trust. More broadly, it suggests that text-generation benchmarks should separate semantic readability from glyph-level correctness. Practical deployment will still require testing across fonts, artistic styles, complex backgrounds, and other writing systems, but TextVerifier offers a concrete framework for making that distinction measurable.

Source: Hugging Face Daily Papers